IMPROVING THE DATA-DRIVEN SPARSE SENSOR PLACEMENT METHOD FOR LARGE-DIMENSIONAL DYNAMICAL SYSTEMS

Abstract

 While observations play essential roles in reducing forecast errors in numerical weather forecasts, optimization methods of sensor placements have not been developed yet. Recently, a data-driven sparse sensor placement (SSP) method based on the proper orthogonal decomposition (POD) was proposed for determining efficient sensor placements as well as reconstructing spatial fields of dynamical systems. The SSP method identifies observation placements by maximizing Fisher information matrix F determined by the POD modes. This study aims at improving the SSP methods for large-dimensional dynamical systems such as atmosphere. Through a series of experiments using sea surface temperature (SST) data, reconstruction accuracy was successfully improved using data assimilation approach relative to the original reconstruction method of the SSP. We found that observation placements using singular values is more sensitive to primary modes, and more suitable when more modes are used for the SSP. Minimizing the trace of F⁻¹ yielded smaller errors in reconstructed fields than maximizing the determinant of F, suggesting that the former is more suitable implementation for maximizing Fisher information matrix F for the problem of SST.